improve build support for native arm targets and multi-arch build

[raghavduddala]

Motivation

• Building cuVSLAM natively on Jetson required several undocumented steps and workarounds.
• The hardcoded -arch=all bypassed CMake's architecture management, and there was no way to target a specific GPU from the build script.
• The speedup test filter bug caused false test failures.

Summary of Changes

• Replaced hardcoded -arch=all in libs/cuda_modules/cuda_kernels/CMakeLists.txt with CMake's native
  CMAKE_CUDA_ARCHITECTURES via the CUDA_ARCHITECTURES target property
• Add --cuda_arch flag to build_release.sh for targeting specific GPU architectures (e.g. --cuda_arch=87 for Jetson Orin Nano) - Add default CMAKE_CUDA_ARCHITECTURES=all fallback in top-level CMakeLists.txt
• Fix GTEST_FILTER in build_release.sh to exclude both SpeedUp and Speedup test naming conventions — the case-sensitive filter was missing the image_pyramid_test speedup tests
• Add "Build natively on Jetson" section to README documenting required dev packages (libcublas-dev, libcusolver-dev), Git LFS setup, and --cuda_arch usage

---

Tested the following:

• Built with --cuda_arch=87 on Jetson Orin Nano (CUDA 12.6) with all module tests passing
  ./build_releases.sh --cuda_arch=87 --modules_test --build_lib

• Built with --cuda_arch=86 on x86_64 (Ubuntu 22.04, nvidia-driver: 580, CUDA 13, RTX 3070)
  ./build_releases.sh --cuda_arch=87 --modules_test --build_lib

• Built without --cuda_arch on x86_64 (Ubuntu 22.04, nvidia-driver: 580, CUDA 13, RTX 3070) which builds for all architectures (SM_75 through SM_121), and all the module tests passed
  ./build_releases.sh --modules_test --build_lib

• Verified cuobjdump -lelf shows only SM_86 with --cuda_arch=86 and all supported architectures without the flag

  Closes Building cuVSLAM for multiple archs and Module Tests #39